Skip to contentSkip to conversation

mm: limit filemap_fault readahead to VMA boundaries

PATCH Superseded Frederick Mayle 4 participants 7 messages Apr 22 – Apr 23

Subsystems (3)

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 mm: limit filemap_fault readahead to VMA boundaries

When a file mapping covers a strict subset of a file, an access to the mapping can trigger readahead of file pages outside the mapped region. Readahead is meant to prefetch pages likely to be accessed soon, but these pages aren't accessible via the same means, so it fair to say we don't have a good indicator they'll be accessed soon. Take an ELF file for example: An access to the end of a program's read-only segment isn't a sign that nearby file contents will be accessed next (they are likely to be mapped discontiguously, or not at all). The pressure from loading these pages into the cache can evict more useful pages.

To improve the behavior, make three changes:

* Introduce a new readahead_control option, max_index, as a hard limit on the readahead. The existing file_ra_state->size can't be used as a limit, it is more of a hint and can be increased by various heuristics.
* Set readahead_control->max_index to the end of the VMA in all of the readahead paths that can be triggered from a fault on a file mapping (both "sync" and "async" readahead).
* Limit the read-around range start to the VMA's start.

Note that these changes only affect readahead triggered in the context of a fault, they do not affect readahead triggered by read syscalls. If a user mixes the two types of accesses, the behavior is expected to be the following: if a fault causes readahead and places a PG_readahead marker and then a read(2) syscall hits the PG_readahead marker, the resulting async readahead *will not* be limited to the VMA end. Conversely, if a read(2) syscall places a PG_readahead marker and then a fault hits the marker, the async readahead *will* be limited to the VMA end.

There is an edge case that the above motivation glosses over: A single file mapping might be backed by multiple VMAs. For example, a whole file could be mapped RW, then part of the mapping made RO using mprotect. This patch would hurt performance of a sequential read of such a mapping, the degree depending on how fragmented the VMAs are. A usage pattern like that is likely rare and already suffering from sub-optimal performance because, e.g., the fragmented VMAs limit the fault-around, so each VMA boundary in a sequential read would cause a minor fault. Still, this would make it worse. See a previous discussion of this topic at [1].

Tested by mapping and reading a small subset of a large file, then using the cachestat syscall to verify the number of cached pages didn't exceed the mapping size.

In practical scenarios, the effect depends on the specific file and usage. Sometimes there is no effect at all, but, for some ELF files in Android, we see ~20% fewer pages pull into the cache.

A comprehensive performance evaluation hasn't been done, but, in addition to the anecdontal memory savings mentioned above, a benchmark was run with fio 3.38, showing neutral looking results:

    /data/local/tmp/fio --version

    fio --name=mmap_test --ioengine=mmap --rw=read --bs=4k \
        --offset=1G --size=1G --filesize=3G --numjobs=1 \
        --filename=testfile.bin

        Before: 4366.6 MiB/s (avg of 3459, 4592, 4613, 4697, 4472)
        After:  4444.0 MiB/s (avg of 4633, 4655, 4511, 4571, 3850)
                +1.7%

    Same, with --ioengine=mmap --rw=randread

        Before: 445.6 MiB/s  (avg of 446, 447, 442, 452, 441)
        After:  447.0 MiB/s  (avg of 447, 446, 446, 451, 445)
                +0.3%

    Same, with --ioengine=psync --rw=read

        Before: 3086.6 MiB/s (avg of 3122, 3094, 3066, 3094, 3057)
        After:  3084.6 MiB/s (avg of 3039, 3103, 3103, 3084, 3094)
                -0.06%

    Same, with --ioengine=psync --rw=randread

        Before: 2226.4 MiB/s (avg of 2256, 2183, 2207, 2265, 2221)
        After:  2231.4 MiB/s (avg of 2236, 2241, 2236, 2193, 2251)
                +0.2%

[1] https://lore.kernel.org/all/ivnv2crd3et76p2nx7oszuqhzzah756oecn5yuykzqfkqzoygw@yvnlkhjjssoz/

Cc Andrew Morton <[email protected]>
Cc David Hildenbrand <[email protected]>
Cc Jan Kara <[email protected]>
Cc Kalesh Singh <[email protected]>
Cc Lorenzo Stoakes <[email protected]>
Cc Matthew Wilcox <[email protected]>
Cc Suren Baghdasaryan <[email protected]>
---
 include/linux/pagemap.h | 2 ++
 mm/filemap.c            | 4 ++++
 mm/readahead.c          | 5 ++++-
 3 files changed, 10 insertions(+), 1 deletion(-)
include/linux/pagemap.h +2 -0
index ec442af3f886..cc628050bc5e 100644
@@ -1366,6 +1366,7 @@ struct readahead_control {
 	bool dropbehind;
 	bool _workingset;
 	unsigned long _pflags;
+	unsigned long max_index; /* limit readahead to i<=max_index */
 };
 
 #define DEFINE_READAHEAD(ractl, f, r, m, i)				\
@@ -1374,6 +1375,7 @@ struct readahead_control {
 		.mapping = m,						\
 		.ra = r,						\
 		._index = i,						\
+		.max_index = ULONG_MAX,					\
 	}
 
 #define VM_READAHEAD_PAGES	(SZ_128K / PAGE_SIZE)
mm/filemap.c +4 -0
index 4e636647100c..d2f6bef12f58 100644
@@ -3314,6 +3314,8 @@ static struct file *do_sync_mmap_readahead(struct vm_fault *vmf)
 	bool force_thp_readahead = false;
 	unsigned short mmap_miss;
 
+	ractl.max_index = vmf->vma->vm_pgoff + vma_pages(vmf->vma) - 1;
+
 	/* Use the readahead code, even if readahead is disabled */
 	if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
 	    (vm_flags & VM_HUGEPAGE) && HPAGE_PMD_ORDER <= MAX_PAGECACHE_ORDER)
@@ -3396,6 +3398,7 @@ static struct file *do_sync_mmap_readahead(struct vm_fault *vmf)
 		 * mmap read-around
 		 */
 		ra->start = max_t(long, 0, vmf->pgoff - ra->ra_pages / 2);
+		ra->start = max(ra->start, vmf->vma->vm_pgoff);
 		ra->size = ra->ra_pages;
 		ra->async_size = ra->ra_pages / 4;
 		ra->order = 0;
@@ -3438,6 +3441,7 @@ static struct file *do_async_mmap_readahead(struct vm_fault *vmf,
 	}
 
 	if (folio_test_readahead(folio)) {
+		ractl.max_index = vmf->vma->vm_pgoff + vma_pages(vmf->vma) - 1;
 		fpin = maybe_unlock_mmap_for_io(vmf, fpin);
 		page_cache_async_ra(&ractl, folio, ra->ra_pages);
 	}
mm/readahead.c +4 -1
index 7b05082c89ea..95a424b2f3a3 100644
@@ -324,6 +324,8 @@ static void do_page_cache_ra(struct readahead_control *ractl,
 		return;
 
 	end_index = (isize - 1) >> PAGE_SHIFT;
+	if (end_index > ractl->max_index)
+		end_index = ractl->max_index;
 	if (index > end_index)
 		return;
 	/* Don't read past the page containing the last byte of the file */
@@ -471,7 +473,8 @@ void page_cache_ra_order(struct readahead_control *ractl,
 	pgoff_t start = readahead_index(ractl);
 	pgoff_t index = start;
 	unsigned int min_order = mapping_min_folio_order(mapping);
-	pgoff_t limit = (i_size_read(mapping->host) - 1) >> PAGE_SHIFT;
+	pgoff_t limit = min_t(pgoff_t, (i_size_read(mapping->host) - 1) >> PAGE_SHIFT,
+				       ractl->max_index);
 	pgoff_t mark = index + ra->size - ra->async_size;
 	unsigned int nofs;
 	int err = 0;

base-commit: db2a1695b2b6feb071b47b72e61d0359bf1524bf

Signature · 1 line
2.54.0.rc1.555.g9c883467ad-goog
maintainer Reviewed-by on v1 1/1
#1 lore raw
Quoted 84 lines · Frederick Mayle

When a file mapping covers a strict subset of a file, an access to the mapping can trigger readahead of file pages outside the mapped region. Readahead is meant to prefetch pages likely to be accessed soon, but these pages aren't accessible via the same means, so it fair to say we don't have a good indicator they'll be accessed soon. Take an ELF file for example: An access to the end of a program's read-only segment isn't a sign that nearby file contents will be accessed next (they are likely to be mapped discontiguously, or not at all). The pressure from loading these pages into the cache can evict more useful pages.

To improve the behavior, make three changes:

* Introduce a new readahead_control option, max_index, as a hard limit on the readahead. The existing file_ra_state->size can't be used as a limit, it is more of a hint and can be increased by various heuristics.
* Set readahead_control->max_index to the end of the VMA in all of the readahead paths that can be triggered from a fault on a file mapping (both "sync" and "async" readahead).
* Limit the read-around range start to the VMA's start.

Note that these changes only affect readahead triggered in the context of a fault, they do not affect readahead triggered by read syscalls. If a user mixes the two types of accesses, the behavior is expected to be the following: if a fault causes readahead and places a PG_readahead marker and then a read(2) syscall hits the PG_readahead marker, the resulting async readahead *will not* be limited to the VMA end. Conversely, if a read(2) syscall places a PG_readahead marker and then a fault hits the marker, the async readahead *will* be limited to the VMA end.

There is an edge case that the above motivation glosses over: A single file mapping might be backed by multiple VMAs. For example, a whole file could be mapped RW, then part of the mapping made RO using mprotect. This patch would hurt performance of a sequential read of such a mapping, the degree depending on how fragmented the VMAs are. A usage pattern like that is likely rare and already suffering from sub-optimal performance because, e.g., the fragmented VMAs limit the fault-around, so each VMA boundary in a sequential read would cause a minor fault. Still, this would make it worse. See a previous discussion of this topic at [1].

Tested by mapping and reading a small subset of a large file, then using the cachestat syscall to verify the number of cached pages didn't exceed the mapping size.

In practical scenarios, the effect depends on the specific file and usage. Sometimes there is no effect at all, but, for some ELF files in Android, we see ~20% fewer pages pull into the cache.

A comprehensive performance evaluation hasn't been done, but, in addition to the anecdontal memory savings mentioned above, a benchmark was run with fio 3.38, showing neutral looking results:

    /data/local/tmp/fio --version

    fio --name=mmap_test --ioengine=mmap --rw=read --bs=4k \
        --offset=1G --size=1G --filesize=3G --numjobs=1 \
        --filename=testfile.bin

        Before: 4366.6 MiB/s (avg of 3459, 4592, 4613, 4697, 4472)
        After:  4444.0 MiB/s (avg of 4633, 4655, 4511, 4571, 3850)
                +1.7%

    Same, with --ioengine=mmap --rw=randread

        Before: 445.6 MiB/s  (avg of 446, 447, 442, 452, 441)
        After:  447.0 MiB/s  (avg of 447, 446, 446, 451, 445)
                +0.3%

    Same, with --ioengine=psync --rw=read

        Before: 3086.6 MiB/s (avg of 3122, 3094, 3066, 3094, 3057)
        After:  3084.6 MiB/s (avg of 3039, 3103, 3103, 3084, 3094)
                -0.06%

    Same, with --ioengine=psync --rw=randread

        Before: 2226.4 MiB/s (avg of 2256, 2183, 2207, 2265, 2221)
        After:  2231.4 MiB/s (avg of 2236, 2241, 2236, 2193, 2251)
                +0.2%

[1] https://lore.kernel.org/all/ivnv2crd3et76p2nx7oszuqhzzah756oecn5yuykzqfkqzoygw@yvnlkhjjssoz/

Quoted 10 lines
Cc Andrew Morton <[email protected]>
Cc David Hildenbrand <[email protected]>
Cc Jan Kara <[email protected]>
Cc Kalesh Singh <[email protected]>
Cc Lorenzo Stoakes <[email protected]>
Cc Matthew Wilcox <[email protected]>
Cc Suren Baghdasaryan <[email protected]>

Looks good to me. Thanks! Feel free to add:

Honza

Quoted 83 lines
---
 include/linux/pagemap.h | 2 ++
 mm/filemap.c            | 4 ++++
 mm/readahead.c          | 5 ++++-
 3 files changed, 10 insertions(+), 1 deletion(-)
include/linux/pagemap.h +2 -0
index ec442af3f886..cc628050bc5e 100644
@@ -1366,6 +1366,7 @@ struct readahead_control {
 	bool dropbehind;
 	bool _workingset;
 	unsigned long _pflags;
+	unsigned long max_index; /* limit readahead to i<=max_index */
 };
 
 #define DEFINE_READAHEAD(ractl, f, r, m, i)				\
@@ -1374,6 +1375,7 @@ struct readahead_control {
 		.mapping = m,						\
 		.ra = r,						\
 		._index = i,						\
+		.max_index = ULONG_MAX,					\
 	}
 
 #define VM_READAHEAD_PAGES	(SZ_128K / PAGE_SIZE)
mm/filemap.c +4 -0
index 4e636647100c..d2f6bef12f58 100644
@@ -3314,6 +3314,8 @@ static struct file *do_sync_mmap_readahead(struct vm_fault *vmf)
 	bool force_thp_readahead = false;
 	unsigned short mmap_miss;
 
+	ractl.max_index = vmf->vma->vm_pgoff + vma_pages(vmf->vma) - 1;
+
 	/* Use the readahead code, even if readahead is disabled */
 	if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
 	    (vm_flags & VM_HUGEPAGE) && HPAGE_PMD_ORDER <= MAX_PAGECACHE_ORDER)
@@ -3396,6 +3398,7 @@ static struct file *do_sync_mmap_readahead(struct vm_fault *vmf)
 		 * mmap read-around
 		 */
 		ra->start = max_t(long, 0, vmf->pgoff - ra->ra_pages / 2);
+		ra->start = max(ra->start, vmf->vma->vm_pgoff);
 		ra->size = ra->ra_pages;
 		ra->async_size = ra->ra_pages / 4;
 		ra->order = 0;
@@ -3438,6 +3441,7 @@ static struct file *do_async_mmap_readahead(struct vm_fault *vmf,
 	}
 
 	if (folio_test_readahead(folio)) {
+		ractl.max_index = vmf->vma->vm_pgoff + vma_pages(vmf->vma) - 1;
 		fpin = maybe_unlock_mmap_for_io(vmf, fpin);
 		page_cache_async_ra(&ractl, folio, ra->ra_pages);
 	}
mm/readahead.c +4 -1
index 7b05082c89ea..95a424b2f3a3 100644
@@ -324,6 +324,8 @@ static void do_page_cache_ra(struct readahead_control *ractl,
 		return;
 
 	end_index = (isize - 1) >> PAGE_SHIFT;
+	if (end_index > ractl->max_index)
+		end_index = ractl->max_index;
 	if (index > end_index)
 		return;
 	/* Don't read past the page containing the last byte of the file */
@@ -471,7 +473,8 @@ void page_cache_ra_order(struct readahead_control *ractl,
 	pgoff_t start = readahead_index(ractl);
 	pgoff_t index = start;
 	unsigned int min_order = mapping_min_folio_order(mapping);
-	pgoff_t limit = (i_size_read(mapping->host) - 1) >> PAGE_SHIFT;
+	pgoff_t limit = min_t(pgoff_t, (i_size_read(mapping->host) - 1) >> PAGE_SHIFT,
+				       ractl->max_index);
 	pgoff_t mark = index + ra->size - ra->async_size;
 	unsigned int nofs;
 	int err = 0;

base-commit: db2a1695b2b6feb071b47b72e61d0359bf1524bf

Signature · 1 line
2.54.0.rc1.555.g9c883467ad-goog
Signature · 2 lines
Jan Kara <[email protected]>
SUSE Labs, CR
maintainer
#2 lore raw
In reply to Frederick Mayle
Quoted 9 lines · Frederick Mayle

When a file mapping covers a strict subset of a file, an access to the mapping can trigger readahead of file pages outside the mapped region. Readahead is meant to prefetch pages likely to be accessed soon, but these pages aren't accessible via the same means, so it fair to say we don't have a good indicator they'll be accessed soon. Take an ELF file for example: An access to the end of a program's read-only segment isn't a sign that nearby file contents will be accessed next (they are likely to be mapped discontiguously, or not at all). The pressure from loading these pages into the cache can evict more useful pages.

The problem is that we might (for example) use mprotect() to mark a portion of the file as being unmodifiable, but nevertheless still want to prefetch through it (since it will be read, just not written). I'm sure this solves your problem, but I'm not sure it covers all use cases.

maintainer
#3 lore raw
Quoted 11 lines · Matthew Wilcox
Quoted 9 lines · Frederick Mayle

When a file mapping covers a strict subset of a file, an access to the mapping can trigger readahead of file pages outside the mapped region. Readahead is meant to prefetch pages likely to be accessed soon, but these pages aren't accessible via the same means, so it fair to say we don't have a good indicator they'll be accessed soon. Take an ELF file for example: An access to the end of a program's read-only segment isn't a sign that nearby file contents will be accessed next (they are likely to be mapped discontiguously, or not at all). The pressure from loading these pages into the cache can evict more useful pages.

Quoted 4 lines

The problem is that we might (for example) use mprotect() to mark a portion of the file as being unmodifiable, but nevertheless still want to prefetch through it (since it will be read, just not written). I'm sure this solves your problem, but I'm not sure it covers all use cases.

Well, I'm not sure whether all the usecases are covered either but is what you describe above something you'd expect people to commonly do? In general sequential reading through mmap seems to be already relatively rare...

Honza

Signature · 2 lines
Jan Kara <[email protected]>
SUSE Labs, CR
#4 lore raw
Quoted 10 lines · Jan Kara
Quoted 4 lines · Matthew Wilcox

The problem is that we might (for example) use mprotect() to mark a portion of the file as being unmodifiable, but nevertheless still want to prefetch through it (since it will be read, just not written). I'm sure this solves your problem, but I'm not sure it covers all use cases.

Well, I'm not sure whether all the usecases are covered either but is what you describe above something you'd expect people to commonly do? In general sequential reading through mmap seems to be already relatively rare...

I'm curious if anyone knows of a use case like this. I'd think that, if a user is making heavy usage of mprotect, they'd be advance enough to be a step away (or a few steps beyond) wanting to manually trigger readahead with madvise or similar.

In recent history, commit 38b0ece6d763 ("mm/filemap: allow arch to request folio size for exec memory") disabled readahead for VM_EXEC mappings and there doesn't seem to have been pushback yet. Of course, non-exec mappings almost certainly have more varied uses.

#5 lore raw
In reply to Frederick Mayle
Quoted 42 lines · Frederick Mayle

When a file mapping covers a strict subset of a file, an access to the mapping can trigger readahead of file pages outside the mapped region. Readahead is meant to prefetch pages likely to be accessed soon, but these pages aren't accessible via the same means, so it fair to say we don't have a good indicator they'll be accessed soon. Take an ELF file for example: An access to the end of a program's read-only segment isn't a sign that nearby file contents will be accessed next (they are likely to be mapped discontiguously, or not at all). The pressure from loading these pages into the cache can evict more useful pages.

To improve the behavior, make three changes:

* Introduce a new readahead_control option, max_index, as a hard limit on the readahead. The existing file_ra_state->size can't be used as a limit, it is more of a hint and can be increased by various heuristics.
* Set readahead_control->max_index to the end of the VMA in all of the readahead paths that can be triggered from a fault on a file mapping (both "sync" and "async" readahead).
* Limit the read-around range start to the VMA's start.

Note that these changes only affect readahead triggered in the context of a fault, they do not affect readahead triggered by read syscalls. If a user mixes the two types of accesses, the behavior is expected to be the following: if a fault causes readahead and places a PG_readahead marker and then a read(2) syscall hits the PG_readahead marker, the resulting async readahead *will not* be limited to the VMA end. Conversely, if a read(2) syscall places a PG_readahead marker and then a fault hits the marker, the async readahead *will* be limited to the VMA end.

There is an edge case that the above motivation glosses over: A single file mapping might be backed by multiple VMAs. For example, a whole file could be mapped RW, then part of the mapping made RO using mprotect. This patch would hurt performance of a sequential read of such a mapping, the degree depending on how fragmented the VMAs are. A usage pattern like that is likely rare and already suffering from sub-optimal performance because, e.g., the fragmented VMAs limit the fault-around, so each VMA boundary in a sequential read would cause a minor fault. Still, this would make it worse. See a previous discussion of this topic at [1].

Quoted 7 lines

Tested by mapping and reading a small subset of a large file, then using the cachestat syscall to verify the number of cached pages didn't exceed the mapping size.

In practical scenarios, the effect depends on the specific file and usage. Sometimes there is no effect at all, but, for some ELF files in Android, we see ~20% fewer pages pull into the cache.

Didn't Android have a gigantically modified RA window? Could this be why you're seeing such large effects? Or is this no longer the case?

Signature · 1 line
Pedro
#6 lore raw
Quoted 48 lines · Pedro Falcato
Quoted 46 lines · Frederick Mayle

When a file mapping covers a strict subset of a file, an access to the mapping can trigger readahead of file pages outside the mapped region. Readahead is meant to prefetch pages likely to be accessed soon, but these pages aren't accessible via the same means, so it fair to say we don't have a good indicator they'll be accessed soon. Take an ELF file for example: An access to the end of a program's read-only segment isn't a sign that nearby file contents will be accessed next (they are likely to be mapped discontiguously, or not at all). The pressure from loading these pages into the cache can evict more useful pages.

To improve the behavior, make three changes:

* Introduce a new readahead_control option, max_index, as a hard limit on the readahead. The existing file_ra_state->size can't be used as a limit, it is more of a hint and can be increased by various heuristics.
* Set readahead_control->max_index to the end of the VMA in all of the readahead paths that can be triggered from a fault on a file mapping (both "sync" and "async" readahead).
* Limit the read-around range start to the VMA's start.

Note that these changes only affect readahead triggered in the context of a fault, they do not affect readahead triggered by read syscalls. If a user mixes the two types of accesses, the behavior is expected to be the following: if a fault causes readahead and places a PG_readahead marker and then a read(2) syscall hits the PG_readahead marker, the resulting async readahead *will not* be limited to the VMA end. Conversely, if a read(2) syscall places a PG_readahead marker and then a fault hits the marker, the async readahead *will* be limited to the VMA end.

There is an edge case that the above motivation glosses over: A single file mapping might be backed by multiple VMAs. For example, a whole file could be mapped RW, then part of the mapping made RO using mprotect. This patch would hurt performance of a sequential read of such a mapping, the degree depending on how fragmented the VMAs are. A usage pattern like that is likely rare and already suffering from sub-optimal performance because, e.g., the fragmented VMAs limit the fault-around, so each VMA boundary in a sequential read would cause a minor fault. Still, this would make it worse. See a previous discussion of this topic at [1].

Tested by mapping and reading a small subset of a large file, then using the cachestat syscall to verify the number of cached pages didn't exceed the mapping size.

Quoted 6 lines
Quoted 3 lines

In practical scenarios, the effect depends on the specific file and usage. Sometimes there is no effect at all, but, for some ELF files in Android, we see ~20% fewer pages pull into the cache.

Didn't Android have a gigantically modified RA window? Could this be why you're seeing such large effects? Or is this no longer the case?

On the device I used to test this, the relevant storage device had a readahead size of 128kb. In general, it can be configured by device OEMs, so perhaps some devices in the ecosystem have giant RA windows.

Android binaries can have a lot of padding and sometimes unrelated APK sections adjacent in the same file (see https://lwn.net/Articles/1016860/). Those may be the source of the effect, but I haven't verified.

AppImage files have some similarities. They are an ELF files that include a mountable filesystem. Here is a quick test I ran on Debian with the latest neovim appimage (./cachestat is a binary that simply invokes the cachestat syscall on a file):

    echo 3 >/proc/sys/vm/drop_caches && \
        ./cachestat ~/nvim-linux-x86_64.appimage && \
        ~/nvim-linux-x86_64.appimage -es >/dev/null && \
        ./cachestat ~/nvim-linux-x86_64.appimage

This patch reduces nr_cache from 4134 to 2131.

Quoted 3 lines

--
Pedro